← Methods & evidenceHIC · per-arm model card
Version gdpa1-hic-research-v1
No confident prediction · head disabledRecorded capped PROPHET-Ab HIC retention time in minutes for monospecific IgGs under the source protocol; no assembled-bispecific transfer.
Training and data
Ginkgo GDPa1 v1.2: 246 source records, 242 HIC measurements; original folds with additional paired-homology purging.
Production per-arm models fitted: no. Research fits completed: yes. Research development records: 140. Research fits and single-study held-out metrics exist. Independent study transfer, complete weight provenance and product applicability gates remain unmet. Candidate numerical predictions abstain.
Release checks
- homology Purging: met in the documented research experiment
- nearest Neighbor Baseline: met in the documented research experiment
- descriptor And Composition Baselines: met in the documented research experiment
- heldout Calibration Evaluated: met in the documented research experiment
- external Comparable Study: not met for production promotion
- full Weight Provenance: not met for production promotion
- bispecific Applicability: not met for production promotion
- commercial Usage Review Complete: not met for production promotion
Metrics and coverage
Research benchmark · recorded capped HIC retention time
Original GDPa1 monospecific IgG assay, in minutes. This evaluates software models on experimental labels; it supplies no experimental measurement or prediction for your construct. Fixed development, calibration and test splits were frozen before model fitting. Source protocol and study transfer remain limitations.
Development: 140 records in 138 homology groups. Calibration: 48 records in 47 groups. Test: 53 records in 52 groups. These are counts, not independent biological validation claims.
| Research model | Test MAE (min), 95% interval | Test Spearman rho, 95% interval |
|---|---|---|
| Training-mean baseline | 0.232 [0.182, 0.282] | Not defined |
| Sequence-neighbor baseline | 0.349 [0.272, 0.436] | -0.024 [-0.312, 0.269] |
| AA composition + ridge | 0.187 [0.133, 0.241] | 0.502 [0.271, 0.679] |
| Descriptor ridge | 0.214 [0.171, 0.261] | 0.485 [0.237, 0.702] |
| Descriptor random forest | 0.233 [0.180, 0.286] | 0.329 [0.067, 0.566] |
| Provided ESM-2 8M + ridge | 0.206 [0.149, 0.264] | 0.412 [0.137, 0.661] |
Development-only selection chose Descriptor ridge. No model was reselected using the final test results. Intervals use 1000 homology-group bootstrap replicates; they do not guarantee performance in another study or architecture. Provided ESM-2 8M features are a separate provenance-limited research baseline, never Module J scores. Original experimental study · Model card and promotion gates.
Stricter one-chain sensitivity · explicitly post-hoc
The original paired-chain split still allowed 2 test VH and 16 test VL sequences at ≥90% development identity. Grouping antibodies sharing either chain retains 84 development records in 75 groups, 34 calibration records in 28 groups, and the same 53 test records in 26 groups. This diagnostic was frozen after the first results; it is not an independent replication or a new model-selection opportunity.
| Fixed model · alpha 100 for ridge | Test MAE (min), 95% group interval | Whole-group coverage · nominal 90% |
|---|---|---|
| Training-mean baseline | 0.230 [0.183, 0.270] | 92.3% of 26 groups; descriptive panel fraction |
| AA composition + ridge | 0.205 [0.154, 0.250] | 92.3% of 26 groups; descriptive panel fraction |
| Descriptor ridge | 0.223 [0.180, 0.263] | 92.3% of 26 groups; descriptive panel fraction |
| Provided ESM-2 8M + ridge | 0.231 [0.182, 0.278] | 80.8% of 26 groups; descriptive panel fraction |
The provided ESM baseline covers only 80.8% of groups at nominal 90%; it does not pass that empirical coverage check. Selected descriptor ridge remains the original development-selected model; no product predictor is promoted. Shared germline and cross-study applicability remain unestablished.
Do not use for
Format selection, assembled-bispecific developability, affinity, potency, clinical decisions or experimental success claims. The app emits abstentions with null numerical fields for this head.
Existing research experiments and measured retrospective results